Add Windows CUDA llama.cpp release asset

.github/workflows/build-llamacpp-binaries.yml

@@ -82,22 +82,72 @@ jobs:
               -DGGML_FMA=ON
               -DGGML_F16C=ON
               -DGGML_BMI2=ON
+          - os: windows-latest
+            name: windows-x64-cuda
+            cuda: true
+            cuda_version: '13.2.0'
+            cuda_architectures: '86'
+            build_target: llama-funasr-sensevoice
+            timeout_minutes: 90
+            cmake_flags: >-
+              -DGGML_NATIVE=OFF
+              -DGGML_CUDA=ON
+              -DGGML_CUDA_FORCE_CUBLAS=ON
+              -DGGML_CUDA_FA=OFF
+              -DGGML_CUDA_NCCL=OFF
+              -DGGML_CUDA_GRAPHS=OFF
+              -DGGML_AVX=OFF
+              -DGGML_AVX2=OFF
+              -DGGML_AVX_VNNI=OFF
+              -DGGML_AVX512=OFF
+              -DGGML_AVX512_VBMI=OFF
+              -DGGML_AVX512_VNNI=OFF
+              -DGGML_AVX512_BF16=OFF
+              -DGGML_FMA=OFF
+              -DGGML_F16C=OFF
+              -DGGML_BMI2=OFF
     runs-on: ${{ matrix.os }}
+    timeout-minutes: ${{ matrix.timeout_minutes || 30 }}
     defaults:
       run:
         shell: bash
         working-directory: runtime/llama.cpp
     steps:
       - uses: actions/checkout@v4
+      - name: Install CUDA Toolkit
+        if: matrix.cuda
+        uses: Jimver/cuda-toolkit@v0.2.35
+        with:
+          cuda: ${{ matrix.cuda_version }}
+          method: network
+          log-file-suffix: ${{ matrix.name }}.txt
+      - name: Show CUDA compiler
+        if: matrix.cuda
+        run: nvcc --version
       - name: Configure and build
         run: |
           # Release binaries must be portable across user machines, not tuned to
           # the ephemeral CI runner CPU. A native ggml build may emit AVX512 or
           # AVX-VNNI instructions and then crash with SIGILL on ordinary CPUs.
           # Keep the default x64 package conservative, and publish explicit
           # x64-avx2 assets for machines that support AVX2/FMA/F16C/BMI2.
-          cmake -B build -DCMAKE_BUILD_TYPE=Release ${{ matrix.cmake_flags }}
-          cmake --build build --config Release -j 2
+          # The CUDA release asset is SenseVoiceSmall-specific for now; building
+          # only that target keeps the Windows CUDA job bounded and matches the
+          # backend currently exposed by `--backend cuda`.
+          cmake_args=(
+            -B build
+            -DCMAKE_BUILD_TYPE=Release
+            ${{ matrix.cmake_flags }}
+          )
+          if [ -n "${{ matrix.cuda_architectures }}" ]; then
+            cmake_args+=("-DCMAKE_CUDA_ARCHITECTURES=${{ matrix.cuda_architectures }}")
+          fi
+          cmake "${cmake_args[@]}"
+          if [ -n "${{ matrix.build_target }}" ]; then
+            cmake --build build --config Release -j 2 --target "${{ matrix.build_target }}"
+          else
+            cmake --build build --config Release -j 2
+          fi
       - name: Package
         run: |
           mkdir -p pkg
@@ -135,4 +185,4 @@ jobs:
           gh release create "$tag" dist/* \
             --repo "${{ github.repository }}" \
             --title "FunASR llama.cpp runtime $version" \
-            --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime: SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD. Download the default quantized model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\` (the helper requires the Hugging Face CLI: \`pip install -U huggingface_hub\`), then run \`llama-funasr-cli\` / \`llama-funasr-sensevoice\` / \`llama-funasr-paraformer\`. Use the default x64 asset for maximum CPU compatibility; use the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2 for higher throughput. No Python ASR runtime or local build is required. Docs: $docs"
+            --notes "Prebuilt self-contained binaries for the FunASR llama.cpp / GGUF runtime: SenseVoice, Paraformer and Fun-ASR-Nano with built-in FSMN-VAD. Download the default quantized model with \`bash download-funasr-model.sh <sensevoice|paraformer|nano>\` (the helper requires the Hugging Face CLI: \`pip install -U huggingface_hub\`), then run \`llama-funasr-cli\` / \`llama-funasr-sensevoice\` / \`llama-funasr-paraformer\`. Use the default x64 asset for maximum CPU compatibility; use the x64-avx2 asset on CPUs with AVX2/FMA/F16C/BMI2 for higher throughput. The Windows CUDA asset is \`windows-x64-cuda\`; it requires an NVIDIA driver compatible with the CUDA Toolkit version configured by the release workflow, targets CUDA architecture 86, and enables SenseVoiceSmall graph execution with \`llama-funasr-sensevoice ... --backend cuda\`. Build from source for other GPU architectures. No Python ASR runtime or local build is required. Docs: $docs"

runtime/llama.cpp/CMakeLists.txt

@@ -24,6 +24,9 @@ if(NOT CMAKE_BUILD_TYPE AND NOT CMAKE_CONFIGURATION_TYPES)
 endif()
 
 include(FetchContent)
+if(POLICY CMP0169)
+    cmake_policy(SET CMP0169 OLD)
+endif()
 set(LLAMA_BUILD_TESTS    OFF CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_EXAMPLES OFF CACHE BOOL "" FORCE)
 set(LLAMA_BUILD_TOOLS    OFF CACHE BOOL "" FORCE)
@@ -32,7 +35,11 @@ set(LLAMA_CURL           OFF CACHE BOOL "" FORCE)
 FetchContent_Declare(llama
     GIT_REPOSITORY https://github.com/ggml-org/llama.cpp.git
     GIT_TAG        8086439a4cea94c71a5dfb8fe4ad1546aebd640f)
-FetchContent_MakeAvailable(llama)
+FetchContent_GetProperties(llama)
+if(NOT llama_POPULATED)
+    FetchContent_Populate(llama)
+    add_subdirectory(${llama_SOURCE_DIR} ${llama_BINARY_DIR})
+endif()
 
 find_package(Threads REQUIRED)
 set(FUNASR_COMMON ${CMAKE_CURRENT_SOURCE_DIR}/funasr-common)

runtime/llama.cpp/README.md

@@ -72,9 +72,13 @@ cmake --build build-cuda -j --target llama-funasr-sensevoice
   -m sensevoice-small-f16.gguf -a sample.wav --backend cuda
 ```
 
-`--backend cpu` remains the default and is what the current cross-platform
-prebuilt binaries use. A binary built without `-DGGML_CUDA=ON` exits with a clear
-message if `--backend cuda` is requested.
+`--backend cpu` remains the default and is what the portable cross-platform
+prebuilt binaries use. Tagged runtime releases also publish a
+`funasr-llamacpp-windows-x64-cuda.zip` package when the Windows CUDA release job
+passes; it requires an NVIDIA driver compatible with the CUDA Toolkit version
+configured by that release job and targets CUDA architecture 86. Build from source
+for other GPU architectures. A binary built without `-DGGML_CUDA=ON` exits with a
+clear message if `--backend cuda` is requested.
 
 ## Build (shared)
 ```bash

runtime/llama.cpp/tests/test_release_workflow.py

@@ -0,0 +1,41 @@
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[3]
+WORKFLOW = ROOT / ".github" / "workflows" / "build-llamacpp-binaries.yml"
+
+
+def test_windows_cuda_release_asset_is_in_matrix():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+
+    assert "name: windows-x64-cuda" in workflow
+    assert "cuda: true" in workflow
+    assert "windows-x64-cuda" in workflow
+    assert "cuda_architectures: '86'" in workflow
+    assert "build_target: llama-funasr-sensevoice" in workflow
+    assert "timeout_minutes: 90" in workflow
+
+
+def test_windows_cuda_build_uses_cuda_toolkit_and_flags():
+    workflow = WORKFLOW.read_text(encoding="utf-8")
+
+    assert "Jimver/cuda-toolkit" in workflow
+    assert "if: matrix.cuda" in workflow
+    assert "-DGGML_CUDA=ON" in workflow
+    assert "-DGGML_CUDA_FORCE_CUBLAS=ON" in workflow
+    assert "-DGGML_CUDA_FA=OFF" in workflow
+    assert "-DGGML_CUDA_NCCL=OFF" in workflow
+    assert "CMAKE_CUDA_ARCHITECTURES=${{ matrix.cuda_architectures }}" in workflow
+    assert "cmake \"${cmake_args[@]}\"" in workflow
+    assert "--target \"${{ matrix.build_target }}\"" in workflow
+
+
+def test_release_notes_explain_cpu_and_cuda_windows_assets():
+    readme = (ROOT / "runtime" / "llama.cpp" / "README.md").read_text(encoding="utf-8")
+
+    assert "windows-x64-cuda" in readme
+    assert "--backend cuda" in readme
+    assert "Windows CUDA" in readme
+    assert "CUDA architecture 86" in readme
+    assert "Build from source" in readme
+    assert "other GPU architectures" in readme